Blog posts

2026

Three Levels of Biological Inspiration: Bridging NeuroAI and Neuromorphic Engineering

8 minute read

Published:

Artificial Intelligence is advancing at ever increasing speed, learning from massive data and running in huge datacenters. Modern AI does not look much like biological intelligence. So what role does neuroscience play in the context of AI? This question was at the center of a recent workshop organized at Cosyne 2026: “Biologically-Inspired Artificial Intelligence”. Here I show my perspective on this intriguing research question, structuring inspiration from biology into three distinct levels: mechanical, system-level, and behavioral. As a neuromorphic engineer, I am used to working at the mechanical level, replicating the biophysics of neurons and synapses with electronics. However, I see great opportunities to work at a higher level of abstraction by drawing inspiration from biological computation. This hints at the unification of NeuroAI with Neuromorphic Computing.

2024

The Memory Technology Landscape

11 minute read

Published:

Memory technology has become the major battleground for integrated chip innovation. Our chips’ integrated memory is ever-growing and now occupies a considerable portion of commercial silicon dies. While Static-Random-Access-Memory technology dominates the on-chip memory market, a recent article pointed out that SRAM’s scaling - fueled so far by Moore’s Law - is hitting a wall. TSMC’s 3nm node only provides a minor 5% improvement in SRAM bitcell density compared to the 5nm node. So what is the future for on-chip memory?

2023

Memristor-Aware-Training for Resilient Neural Networks

10 minute read

Published:

Memristors: a controversial technology with as many supporters as critics. Memristors have attracted considerable attention in the edge-AI industry as a promising memory technology capable of improving energy efficiency and integration density by orders of magnitude compared with standard integrated memory. However, the memristive revolution has been hindered (or delayed) by many technical difficulties. Among these is the problem of variability, i.e., the stochastic behavior memristors exhibit in their conductance during programming. Is it a dealbreaker? This article shows how to address memristive-conductance variability with a particular training methodology inspired by Quantization-Aware-Training, a method used to train heavily quantized networks. The so-called Memristor-Aware-Training introduces memristor-calibrated variability during training so that optimization can adapt to it and converge to a stable weight configuration that tolerates variability.